IAMCCS_Krea2-redhead

First, the name: it has nothing to do with redheads. Nothing in the code, README, or UI does anything redhead-specific - the honest guess is it's leftover branding from whatever prompt the author was testing when he named it. Treat it as noise and read on.

IAMCCS_Krea2-redhead is a model patcher that makes Krea 2 lean harder into your text prompt. You feed it the Krea 2 model, it hands back a patched clone, and every sampling step that clone pushes text conditioning through the model with more force than the stock forward pass does. It's not a filter bypass, not a LoRA, not a training run - it's an enhancer on the model's text-fusion path, and it's a reasonable tool to have when Krea 2 is the model you're on.

Why you might need that: Krea 2 is a CFG-1, natural-language model whose biggest community complaint is that its trained-in alignment flattens things - characters come out expressionless, prompt clauses get silently ignored. The usual remedies are an uncensor LoRA or a conditioning-rebalance node. This node is a different lever entirely: it doesn't touch the refusal gate, it makes whatever the model is willing to do follow your words harder. If your problem is "the model heard me but stayed timid," that's this tool's job.

How it works

Krea 2 merges multi-layer Qwen3-VL text features into the diffusion transformer through a text-fusion block (txtfusion). The node detects that the model has the real Krea 2 layout (12 tap layers at 2560 dim), then wraps the block's forward pass. On each step it:

1. Runs the text-fusion pass normally to get a reference output.
2. Re-runs it on activations scaled by a 24-chunk boost profile, up to a ~15x global multiplier at full strength.
3. Computes the delta between the two passes and adds it back to the reference output - clamped so no token's activations move more than ~75% relative to the reference.

Net effect: text features get amplified without the conditioning blowing out into garbage. It's registered through ComfyUI's patcher_extension wrapper system, so it composes with other model patchers instead of fighting them.

Inputs and outputs

Only four inputs, and you'll realistically touch one:

• model (MODEL) - your Krea 2 Raw or Turbo model node.
• strength (FLOAT, default 1.0, range 0–2) - the one you tune. Start at 1.0; if results look overcooked (blown contrast, the prompt "shouting"), drop to 0.5. 0 disables it cleanly.
• enabled (BOOLEAN, default true) - a soft on/off you can wire up.
• debug (BOOLEAN, default false) - prints per-step RMS/cosine metrics to the console, capped at 8 lines. Turn it on once if you're curious, then turn it off.

Output: model (MODEL), the patched clone. Wire it into your sampler exactly where the stock model would go.

Installing and gotchas

It ships in the same pack as the goyAIcanvas editor, so:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS_goyAIcanvas-easy

Restart ComfyUI. The node has no extra Python dependencies, but it does need a reasonably current ComfyUI - the patcher_extension subsystem it relies on isn't in older builds. On any model that isn't genuine Krea 2 it silently passes through untouched (it self-detects the layout), so if it does nothing, check that you're actually on Krea 2 and not a finetune with a renamed file.

And remember the license split: the node's code is Apache 2.0, but Krea 2's weights sit under the Krea 2 Community License - free commercially below $1M trailing revenue, revocable on 30 days' notice. The node grants you nothing the weights don't already.